Claude Opus 5.5 costs 40% less to run than Claude Opus 5

Near the top of the Hacker News front page, with about eighteen hundred points: Anthropic released Claude Opus five point five, the first model in its new Claude five point five family.

Anthropic says Claude Opus five point five works at the level of Claude Fable five point one on most jobs, and costs forty percent less to run than Claude Opus five. That comes from a lower price per token, and from fewer tokens per task.

Anthropic's own price table, per million tokens. Reading from the prompt cache drops from fifty cents to twenty cents, and Anthropic says cache reads are most of what agentic coding spends. Input drops from five dollars to four, and output from twenty five dollars to twenty.

One early tester finished a six hundred eighty thousand line code migration in less than a day. In Anthropic's own test, translating HAProxy from C into Rust took nine and a half hours, against twelve hours for Claude Fable five point one, at about half the cost.

Commenter sailingparrot notices what the announcement opens with: Anthropic's own recent call to pace the frontier. Everything after that line, they argue, is specific numbers showing the opposite, and the post does open that way, calling this its first release since that call. GPT six Astra has broken an Enigma message from nineteen forty one.

The message went out on the tenth of July, nineteen forty one, and had held out since two thousand five. Astra picked that target itself, wrote its own Enigma simulator and key search code, and cribbed on a repeated place name.

Commenter tantalor pushes on the framing. The write up says the model did it entirely on its own, and in the same breath says it wrote the simulator and the search code, which is where the work went. And a post about reading what you didn't write.

Colin Breck's argument turns on context. You set the constraints and handed over the source and the logs, so you can skim your own output. Your reader has none of that, and has to weigh every line, so they stop.

Commenter zmmmmm reports sending pull requests back for too much description: pages of generated rationale and risk analysis on a twenty line change. Approving it would imply they had read all of it, and there is not time.

FoxDev Studio rebuilds the Visual FoxPro nine language: a compiler and bytecode machine in Rust, compiled to WebAssembly, with forms drawn by React. Of the seventeen hundred elements in the language reference, about fifteen hundred are tested against the original product's own answer.

Commenter mikestew, who says they worked on the Fox team, raises a caveat the site does not cover. A FoxPro database container has no permissions scheme, so it is writable by everyone, and its stored procedures sit in a plain text field.

Brandon Thomas wants a second word. A watermark is visible and claims ownership; a spymark is hidden, and traces your work without your consent. Google's SynthID fits enough in one image for a database identifier plus error correction, and survives re-encoding.

Commenter Retro_Dev calls it steganography under a new name, and offers a defense: assert that what you publish is byte for byte what you made. The open door, they say, is any platform that re-compresses your upload.